Inspired by iSPIRT’s DEPA — the Data Empowerment and Protection Architecture.
Confidential AI Network is the control plane for that job. A data provider publishes a dataset. A model owner writes a Ricardian contract that names the dataset, the purpose, the price, and the clean-room provider. Training starts only after that contract is signed, on a cluster the provider already runs. The signed price is how the data provider is paid for a use they approve, and how the clean-room provider agrees to run that job. The dataset and the model stay with their owners.
Who you are, the contract you sign, and the deployment you are standing in.
Each party, contract, and dataset gets an id shaped like PREFIX-TDP-uuid. The prefix is the deployment. That id is how the API tells a data provider, a model owner, and a clean-room provider apart.
A Ricardian contract is the legal text and the machine record. When the SCITT ledger is enabled, the approval is written there so a later check sees the same event.
The data provider publishes the dataset. The model owner trains. The clean-room provider runs the job. Role checks decide who may sign, start training, or audit.
Each install sets its region, country, jurisdiction, and the frameworks it is configured for. The app shows that card on the homepage. The list is configuration, not a certification.
The running app reads this from its own configuration and shows it on the homepage.
PREFIX-ENTITY-UUID
Entity is TDC, TDP, TSP, CONTRACT, or DATASET. A default install uses the prefix US-EAST, so a data provider looks like US-EAST-TDP-….
That same card lists region, country, jurisdiction, and frameworks such as SOX, FedRAMP, and CCPA. Change the deployment settings and the card changes with them.
A React portal, an Express API, and a DEPA gateway share one Postgres database. Keycloak signs people in on a laptop. Azure uses Entra when you switch the identity provider.
Publishes a dataset with a price and signs only the uses they accept. Their dashboard sums that contract value as revenue. The dataset stays with them. Consent and the privacy budget are the compliance record for that use.
Pays the contract price to train on data they do not receive. Their model key is not stored as raw bytes. The finished job keeps metrics, the artifact, and a provenance leaf on that contract, so they can show which dataset and clean room produced the model.
Signs the contract that names their cluster and the price, and runs the job on capacity they already operate. The kubeconfig stays on their Cloud Provider card. The dataset and the model key are not written onto the contract.
Opens the sealed Merkle tree for a contract and checks a leaf against that root. The check lives in this database. It is not a filing with an outside records office.
One contract, then one training run on a clean room the provider already operates.
The contract is the legal text, the price, and the machine record: dataset, model, purpose, duration, differential-privacy budget when one is set, and the clean room that will run the job.
The data provider signs first. That approves the dataset, the price, and the purpose. The clean-room provider signs next and agrees to run the job. Training stays blocked until both have signed.
Local jobs run Docker, native PyTorch, or MLX on a Mac. Cloud jobs submit to the named AKS, GKE, or OKE cluster. The metrics, the model artifact, and the provenance report stay on that contract.
The contract names the cluster. The kubeconfig path stays on the provider’s Cloud Provider card.
Docker, native, or MLX on the machine that runs the API. ./start-system.sh brings up Keycloak, Postgres, the API, and the portal.
An existing AKS cluster. A missing kubeconfig fails the job. Key Vault lookup fails closed before the job is submitted. A guest report is read only when the confidential node exposes it.
An existing GKE cluster and project. The kubeconfig path is on the provider card or in GCP_CCR_KUBECONFIG.
An existing OKE cluster through the OKE runner. AWS training returns HTTP 501 and does not create a job.
A missing kubeconfig path fails the job. Attestation stays simulated until a SEV-SNP guest report verifies.
The same contract is the place to check signatures, training, and audit.
Who signed, the contract hash, and the status machine are on the contract. Training checks that status before it starts.
Dataset and model key release records the time. The API rejects raw key bytes. Azure attested release returns an encrypted blob and does not decrypt it in the API process.
The model owner can point at this contract for the job, the artifact, and a Merkle leaf. An auditor checks that leaf against the sealed root. The check lives in this system.
The repository is the product. Clone it, run the stack, and follow one contract through training.
Copy config.env and secrets.env, then ./start-system.sh. The API health check is http://localhost:5001/health. The portal is on port 3000.
The portal uses /api on port 5001. DEPA clients use /v1 on port 8000. The gateway translates those calls. It does not run a second product.
The quick start is the shortest path. The product tour shows the screens. Questions go to hello@dpi-apps.space.